Latent-Space Scoring Cuts Video Generation to 1-4 Steps

Featured

01 Video Generation Solves Fast And Good Together

Reward Lightning's core move is optimizing two goals — preference alignment and distillation acceleration — in the same representation space. Normally these run separately. Preference alignment decodes output back to pixels and scores it with a pixel-level reward model. Distillation acceleration works in latent space. The representations don't match, the gradients fight, and fixing one often breaks the other.

The first step trains a reward model (LRM) that scores video directly in latent space, no decode to pixels required. The engineering payoff is real. Video decode is a memory and time sink, and it weighs heavier on the training loop as sampling steps drop and batch size grows. Moving scoring into latent space cuts the most expensive link in the reward chain, freeing budget for few-step training. The second step lets adversarial distillation and preference alignment share that backbone. Both goals read the same latent features and backprop through one network, so the trade-off gets resolved in one place instead of patched serially — accelerate first, align later, with each step cleaning up after the last.

The result: usable output in 1 to 4 steps, a 2.1% VBench average gain, and LRM preference accuracy 11% and 14.7% above pixel-level and latent-level baselines. That second number is the evidence that latent scoring isn't worse than decoding to pixels — the premise the whole skip-the-decode approach rests on. The abstract only reports motivation and these figures, so how cleanly "both at once" splits the trade-off still needs the full paper to confirm.

Key takeaways: - The gradient conflict between preference alignment and distillation is really a representation-space mismatch. Unifying to latent space is worth trying. - Scoring directly in latent space skips the decode-to-pixels cost, valuable for teams doing few-step video generation. - Discount the "have your cake and eat it" claim for now. A 2.1% VBench gain is modest; watch whether the trade-off truly splits under heavy load.

Source: Reward Lightning: Fast Video Generation via Homologous Preference Distillation

02 Safety Alignment Pick The Right Prompt Before You Unlearn

Erasing harmful concepts — gender bias, bioweapons — from a pretrained model stays hard. Existing unlearning methods watch only the weights and ignore prompt shape. Ask a different way and the "erased" concept resurfaces. MPSelectTune changes the angle. Find the worst prompt type — the one that pushes concept prediction accuracy highest — and fine-tune against it specifically. That beats blending every prompt type into one training mix.

The method has two stages. First, fine-tune with a multi-task loss across prompt types. Then minimize concept accuracy on the worst prompt. Across four benchmarks, main-task accuracy rises 2-15% over recent baselines, and worst-case concept accuracy drops up to 17%. For safety teams, this isn't another unlearning algorithm. It's a usable knob: verify robustness along the prompt dimension, because averages hide the backdoor.

Key takeaways: - Prompt-type selection is itself a knob for unlearning, not a side detail of fine-tuning. - Training on the worst prompt holds unlearning robustness better than mixing all prompts. - Evaluate unlearning by worst-case concept accuracy. Averages miss the backdoor that gets around it.

Source: MPSelectTune: Prompt-type Selection for Fine-tuning improves Concept Unlearning in LLMs

03 Multimodal Few-Shot Is Great, Until The Factory Floor

Few-shot generalization dazzles in demos and stalls in factory inspection. Defects are tiny, image resolution is limited, and worse, the text criteria ("no scratches here") don't line up with the pixels the model actually sees. The verdict comes with no defensible basis. GLLS skips fine-tuning entirely. It's training-free, so it needs none of the defect samples you can't collect during early deployment. Instead it runs in-context verification through a dual-stream design. The global logic stream uses SAM 3 to extract checkable visual facts. The local search stream uses Monte Carlo tree search to pick evidence regions within a fixed budget.

Its pitch is verifiability. Every diagnosis traces back to specific visual evidence rather than a model's hunch. If you deploy inspection, the thing to watch isn't a few more points on a few-shot benchmark. It's how much that "traceable evidence" chain delivers on a real line. The abstract reports steady gains over baselines on MMAD-QA and similar datasets, but the practical value of the explainability needs the case studies in the full paper.

Key takeaways: - The training-free route sidesteps the "no defect samples, no fine-tuning" deadlock, fitting cold-start deployment. - The real differentiator isn't the few-shot score. It's verifiability — verdicts that trace to visual evidence. - Whether this explainable chain delivers on a real production line needs concrete cases, not benchmark numbers.

Source: Global Logic and Local Search: Dual-Stream Multimodal In-Context Learning for Verifiable Industrial Anomaly Detection

04 Image Generation Depth And Normals Patch Sparse-View Holes

Reconstructing geometry from a few photos has one glaring flaw: holes and cracks on the surface. Geometry cues are already sparse, and 3DGS (3D Gaussian Splatting) uses discrete points, so it leaves gaps and fractures where the surface should stay continuous. This ECCV work fixes both by filling from the good. It pulls reliable depth from high-confidence regions, then uses a normal prior — surface orientation — to propagate that depth into low-confidence regions and constrain them. An edge regularizer targeting depth discontinuities handles the surface fractures from discrete Gaussians.

On DTU and Tanks-and-Temples, it beats the previous best sparse-view surface reconstruction. If you build 3D assets or scan reconstructions, the point isn't whether another novel-view synthesis demo looks good. It's how much the acknowledged weakness — geometric surfaces — got filled. The exact margin and the behavior on real cluttered scenes still need the full paper and project page.

Key takeaways: - Novel-view synthesis under sparse views is mature. Geometric surface reconstruction is the real weak spot now. - Using high-confidence depth plus a normal prior to constrain low-confidence regions is a practical fill-from-the-good approach. - Teams doing 3D scanning and asset reconstruction should watch geometry-quality metrics, not pretty render demos.

Source: Sparse-View Surface Reconstruction using Gaussian Splatting through High-Confidence Depth Propagation with Normal Priors